Near-ideal model selection by ℓ1 minimization

Abstract

We consider the fundamental problem of estimating the mean of a vector y = Xβ + z, where X is an n × p design matrix in which one can have far more variables than observations, and z is a stochastic error term-the so-called "p > n" setup. When β is sparse, or, more generally, when there is a sparse subset of covariates providing a close approximation to the unknown mean vector, we ask whether or not it is possible to accurately estimate Xβ using a computationally tractable algorithm.We show that, in a surprisingly wide range of situations, the lasso happens to nearly select the best subset of variables. Quantitatively speaking, we prove that solving a simple quadratic program achieves a squared error within a logarithmic factor of the ideal mean squared error that one would achieve with an oracle supplying perfect information about which variables should and should not be included in the model. Interestingly, our results describe the average performance of the lasso; that is, the performance one can expect in an vast majority of cases where Xβ is a sparse or nearly sparse superposition of variables, but not in all cases.Our results are nonasymptotic and widely applicable, since they simply require that pairs of predictor variables are not too collinear.where X is an n × p matrix of explanatory variables, β is a p-dimensional parameter of interest and z = (z 1 , . . . , z n ) is a vector of independent stochastic

Adaptive sensing for sparse recovery

“…solve a convex relaxation of the form 6) for some τ > 0. This 1 -constrained optimization is relatively easy to solve using convex optimization techniques.…”

“…For example, if the columns of A are scaled to have unit norm, recent work [6] suggests that the optimization in (1.6) will succeed (with high probability) only if the magnitudes of the non-zero components exceed a fixed constant times √ log n. In this chapter we will see that this fundamental limitation can again be overcome by sequentially designing the rows of A so that they tend to focus on the relevant components as information is gathered.…”

Dynamic Updating for $\ell_{1}$ Minimization

“…Assume that we have solved (2) using (6) to get the estimatex 0 with our current set of m measurements y. Now suppose that we get one new measurement given as…”

Coherence-based near-oracle performance guarantees for sparse estimation under Gaussian noise

“…Thus, no unbiased estimator can achieve an MSE below (9). A different strategy for appraising estimators is to compare practical techniques with the "oracle" estimator, which is the LS solution among vectors x whose support is Λ 0 [2].…”